Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping
Looped Transformers can generalize well to longer inputs but are unstable because they randomly choose how many times to loop during training, leading to high prediction variance. The work introduces learned stochastic stopping schedules, like RL-Halting, which stabilize predictions and improve the balance between accuracy and stability. This suggests deciding 'when to stop' is a crucial training design choice.